Keep context out of the prompt: RLM-27B beats base Qwen3-27B by 26.9%
dosco · x · 2026-11-22
A TLDR of a new RLM (recursive language model) training paper is making rounds:
- Typical harnesses append full history into one prompt, forcing models to cope with contexts they never trained on;
- RLMs instead keep context outside the prompt, write code over it, and hand smaller pieces to sub-calls — each LLM call stays "locally in-distribution." Trained only on short tasks, this generalized to tasks 2–64x longer and transferred to new domains;
- The key trick is surrogate harness training: the orchestrator is trained on cheap synthetic environments with fake tool calls and fake sub-agents instead of expensive real rollouts;
- RLM-27B beats base Qwen3-27B by 26.9%.
A commenter adds that attention makes models dumber as context grows, and the usual fix is bigger, pricier models — RLM offers another path.
More from coding & agent
- instant_record compiles Rails ActiveRecord models to WASM for instant, offline browser apps — danshipper · 2026-11-22
- Client demoing your AI build internally without prompting is the real adoption signal — aryanXmahajan · 2026-11-22
- Animama: open-source AI desktop companion with VRM avatars and versioned memory files — kavidy · 2026-11-22
- Building an AI team with Claude: an AI employee that already knows its job — blaizedsouza · 2026-11-22
- SteerCamp ships Mac app: delegate work to a crew of AI members on cloud machines — pramodk73 · 2026-11-22
- Open-source Claude Code plugin cuts complete music videos from a single song, now at v1.7 — AIandDesign · 2026-11-22